@ChShersh It was a great documentary, I loved Alexander's personality, he really was the catalyst the team needed at the time, made me watch some of his lectures
People evaluate local AI with:
βCan it solve LeetCode?β
βCan it one-shot a SaaS?β
But i believe the real questions should be:
Does it keep the same JSON schema?
Does it hallucinate fields?
@techczech I run a processing pipeline through gemma4 26b, 31b and 34b, I get consistent structured outputs with thinking switched off. I am running it in production doing 15000 documents a week on premises
@SlimTradeyBaby@OrganicGPT Can you run benchmarks on how it reproduces structured outputs for large documents and have gpt5 rate them? Feed it government contracts for example, have key info that you want extracted then benchmark the models. I find Gemma always beats everythng else, even e4b is consistent
@OrganicGPT Same here, I run structured outputs and it's always returning expected results. I am talking multiple documents through gemma4 26b and 31b, even e4b passes when I run it through my benchmarks, better than the Nemotron 3 Nano 30b and 27b qwen.
@ankkala Processing documents. I've got a pipeline running constantly on my computer fetching government tender documents from multiple sources then organising them to a postgresql server. I use a combination of gemma4 26b, 31b, E2b, E4b. I do about ~15k entries a week
@liquidai Can't wait to try it out. Been hard at work with my document pipelines, currently using the gemma4 31b, but I'd like to swop it for something quicker